思想
we minimize a quadratic binary-assignment function containing two terms:
(1) a unary term assigning query samples to the nearest class prototype, and
(2) a pairwise Laplacian term encouraging nearby query samples to have consistent label assignments.
( 1 ) (1) (1) 每个query要和对应标签的原型距离尽可能近
( 2 ) (2) (2)每个query要尽可能和相似的其他query的标签一样
数学公式
特征提取器 f θ f_{\theta} fθ
使用不同的网络，输出的特征维度也不同
特 征 提 取 器 f θ = { R e s N e t 18 / R e s N e t 50 = 512 d i m e n s i o n M o b i l e N e t = 1024 d i m e n s i o n W R N = 640 d i m e n s i o n D e n s e N e t = 1024 d i m e n s i o n 特征提取器f_{\theta}=\left\{ \begin{aligned} ResNet18/ResNet50 & = & 512 dimension \\ MobileNet & = &1024 dimension \\ WRN & = &640 dimension \\ DenseNet & = &1024 dimension \end{aligned} \right. 特征提取器fθ=⎩⎪⎪⎪⎪⎨⎪⎪⎪⎪⎧ResNet18/ResNet50MobileNetWRNDenseNet====512dimension1024dimension640dimension1024dimension
实现细节
Proposed Algorithm for LaplacianShot
